A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 03:00 UTC
[Misc] Add pytest marker to opt-out of global test cleanup (#3863)
Cade Daniel committed
Apr 4, 2024 at 21:54 UTC
e5043a3e7505aefc59636f7df129dc168da4a7d2
3 files changed
+11
-4
tests/conftest.py
+5
-1
@@ -56,11 +56,15 @@ def cleanup():
56
57
58
@pytest.fixture()
59
-def should_do_global_cleanup_after_test() -> bool:
59
+def should_do_global_cleanup_after_test(request) -> bool:
60
"""Allow subdirectories to skip global cleanup by overriding this fixture.
61
This can provide a ~10x speedup for non-GPU unit tests since they don't need
62
to initialize torch.
63
"""
64
+
65
+ if request.node.get_closest_marker("skip_global_cleanup"):
66
+ return False
67
+
68
return True
69
70
tests/spec_decode/test_batch_expansion.py
+3
@@ -7,6 +7,7 @@ from .utils import create_seq_group_metadata_from_prompts, mock_worker
7
8
9
@pytest.mark.parametrize('num_target_seq_ids', [100])
10
+@pytest.mark.skip_global_cleanup
11
def test_create_target_seq_id_iterator(num_target_seq_ids: int):
12
"""Verify all new sequence ids are greater than all input
13
seq ids.
@@ -27,6 +28,7 @@ def test_create_target_seq_id_iterator(num_target_seq_ids: int):
28
29
30
@pytest.mark.parametrize('k', [1, 2, 6])
31
+@pytest.mark.skip_global_cleanup
32
def test_get_token_ids_to_score(k: int):
33
"""Verify correct tokens are selected for scoring.
34
"""
@@ -53,6 +55,7 @@ def test_get_token_ids_to_score(k: int):
55
56
57
@pytest.mark.parametrize('k', [1, 2, 6])
58
+@pytest.mark.skip_global_cleanup
59
def test_create_single_target_seq_group_metadata(k: int):
60
"""Verify correct creation of a batch-expanded seq group metadata.
61
"""
tests/spec_decode/test_spec_decode_worker.py
+3
-3
@@ -487,7 +487,7 @@ def test_empty_input_batch(k: int, batch_size: int):
487
**execute_model_data.to_dict())
488
489
490
-@torch.inference_mode()
490
+@pytest.mark.skip_global_cleanup
491
def test_init_device():
492
"""Verify SpecDecodeWorker invokes proposer/scorer worker init_device, as
493
well as other GPU initialization.
@@ -537,7 +537,7 @@ def test_init_cache_engine():
537
@pytest.mark.parametrize('available_cpu_blocks', [500])
538
@pytest.mark.parametrize('target_cache_block_size_bytes', [2 * 2 * 4096])
539
@pytest.mark.parametrize('draft_kv_size_bytes', [0, 2 * 2 * 768, 2 * 2 * 4096])
540
-@torch.inference_mode()
540
+@pytest.mark.skip_global_cleanup
541
def test_profile_num_available_blocks(available_gpu_blocks: int,
542
available_cpu_blocks: int,
543
target_cache_block_size_bytes: int,
@@ -584,7 +584,7 @@ def test_profile_num_available_blocks(available_gpu_blocks: int,
584
@pytest.mark.parametrize('target_cache_block_size_bytes',
585
[2 * 2 * 4096, 2 * 2 * 8192])
586
@pytest.mark.parametrize('draft_kv_size_bytes', [0, 2 * 2 * 768, 2 * 2 * 4096])
587
-@torch.inference_mode()
587
+@pytest.mark.skip_global_cleanup
588
def test_split_num_cache_blocks_evenly(available_gpu_blocks: int,
589
target_cache_block_size_bytes: int,
590
draft_kv_size_bytes: int):